I built an open-source CLI that scores any site for AI-agent readiness (0-100)

TL;DR — agent-readiness-cli checks how well your site talks to ChatGPT, Claude, Perplexity and other AI agents. Single-file Python, standard library only, MIT.
Repo: github.com/sspoisk/agent-readiness-cli
Install: pip install agent-readiness-cli

The problem

Every "AI SEO" article in the last six months tells you the same three things:

Add llms.txt
Add JSON-LD with the right @type
Decide what to do about GPTBot, ClaudeBot, PerplexityBot in robots.txt

What none of them give you is a tool that opens your site, looks at it the way a crawler would, and tells you what's actually missing right now.

So I built one.

What it does

$ agent-ready https://example.com
✓ llms.txt               10/15  present, 4.2 KB, 12 URLs
✓ json-ld                23/25  3 block(s), types: Article, Organization, BreadcrumbList
✗ ai-bots-robots.txt      0/20  ClaudeBot, GPTBot disallowed at root
✓ canonical+hreflang     12/15  canonical=set, hreflang langs=['en','ru']
✗ mcp-card                0/10  no /.well-known/mcp.json (optional)
✓ meta                   10/10  10/10 of common signals
✓ sitemap                 5/5   valid, 1250 URLs

  Score: 60 / 100
  Tier: C  (middling — focus on ai-bots-robots.txt, mcp-card)

Enter fullscreen mode Exit fullscreen mode

One number for the whole site, plus seven sub-scores so you know what to fix first.

What gets checked, with weights

Section

Weight

What it looks for

llms.txt

presence, leading H1, at least 3 canonical URLs listed

json-ld

parseable, recognised @type, multiple distinct types

ai-bots-robots.txt

rules for GPTBot / ClaudeBot / Claude-Web / PerplexityBot / Google-Extended / CCBot / Applebot-Extended / Bytespider

canonical+hreflang

self-canonical, reciprocity, x-default for multi-lang

mcp-card

/.well-known/mcp.json valid JSON with name + description

meta

description, og:title, og:description, twitter:card, <html lang>

sitemap

/sitemap.xml exists, valid <urlset> or <sitemapindex>

Bands: A ≥ 90 · B ≥ 75 · C ≥ 55 · D ≥ 35 · F < 35.

I picked these weights based on what I see most often missing in the wild — JSON-LD is heaviest because it's the highest-leverage signal an LLM uses to understand your page kind. Disagree? All weights are in agent_ready/cli.py and contributors are welcome to challenge any of them.

Output formats for different jobs

agent-ready https://example.com           # human summary (default)
agent-ready --full https://example.com    # human summary + every finding
agent-ready --json https://example.com    # machine-readable
agent-ready --csv https://example.com     # one row, append to your monitoring
agent-ready --quiet https://example.com   # just the score; exit code = band

Enter fullscreen mode Exit fullscreen mode

CI gate example

If you want your build to fail when the site drifts below a threshold:

- name: Audit AI-agent readiness
  run: |
    pip install agent-readiness-cli
    SCORE=$(agent-ready --quiet https://your-site.example)
    echo "AI-readiness score: $SCORE"
    [ "$SCORE" -ge 75 ] || { echo "below threshold"; exit 1; }

Enter fullscreen mode Exit fullscreen mode

That single integer-on-stdout is why --quiet exists.

What it does NOT do

It does not crawl your whole site — one URL at a time. If you want the whole site, drive it from a sitemap loop.
It does not fix anything. It tells you what to fix; the fix is on you.
It does not check for vulnerabilities — use OWASP ZAP for that.
It does not validate JSON-LD against the full Schema.org grammar. It checks that types are recognised; for Schema-strict validation, use Google's Rich Results Test.
It does not score Core Web Vitals or accessibility.

If any of those is what you actually need, this is not the right tool.

Why a single file

The whole tool is one Python file plus tests. No third-party dependencies. Standard library only.

A few reasons this matters:

Auditable. You can read every check in one go. No layer of abstraction hides what gets weighted.
Portable. It runs on any box with Python 3.10+. No pip install-and-pray.
No telemetry. It hits your URL only. Nothing else leaves the machine.
Forkable. If you want to add a check or change a weight, fork it. The whole thing is shorter than most config files.

I think the future of small dev tools is a return to this — one file, one job, no surprise dependencies.

Where it sits in the landscape

There are excellent generators (e.g. firecrawl/llmstxt-generator) that produce llms.txt files for you. There are validators for JSON-LD (Google's web tool, schema linters). There are MCP doc tools like langchain-ai/mcpdoc for exposing llms-txt to IDEs.

What didn't exist was the audit slice — a single CLI that opens your URL, looks at the agent-readiness surface end-to-end, and says "score 62, weakest links are X and Y." So I wrote that.

How to try it now

pip install agent-readiness-cli
agent-ready https://your.site

Enter fullscreen mode Exit fullscreen mode

Available on PyPI and GitHub.

If you want continuous monitoring instead of one-off audits, I built it on top of Web-Audit Guardian — the same logic running every 30 min for a domain. The CLI is the audit slice as OSS; the continuous variant is the paid tier. Either way, the methodology is now public.

What's next

A --all mode that follows your sitemap and rolls up to a site-wide score
Optional check for ai.txt (Spawning's draft)
More AI bots if reasonable consensus emerges

Issues, PRs and disagreements with the weights all welcome. The repo is small enough that the bar to contribute is low.

Repo: github.com/sspoisk/agent-readiness-cli
License: MIT
Author: maintained by GuardLabs.

If you run it, let me know what your number is. Always interested to hear which check most surprised people.